3x Faster Search: Parallel Test-Time Scaling With Instructed-Retriever-1
Databricks, Thursday, June 4th, 2026
Databricks announces Instructed-Retriever-1, a retrieval-specialized model that achieves 3x faster search and 2x faster answer generation through parallel test-time scaling.
Databricks has released Instructed-Retriever-1, a retrieval-specialized model that powers their Agent Bricks Knowledge Assistant with significant performance improvements. The model uses parallel test-time scaling to reduce search latency by over 3x and answer generation time by 2x, achieving Time To First Token around two seconds.
Unlike sequential agentic retrieval systems, Instructed-Retriever-1 parallelizes query generation and reranking to improve recall and precision simultaneously while maintaining low latency. The model was trained on synthetic enterprise-style retrieval environments and matches Claude Sonnet 4.5 retrieval quality on benchmarks while using efficient serving optimizations like Mixture-of-Experts architecture and FP8 quantization.